Messages in this thread | | | Date | Mon, 20 Feb 2012 11:46:32 -0800 (PST) | From | Linus Torvalds <> | Subject | [PATCH v2 0/3] More i387 state save/restore work |
| |
This is a slightly updated patch-series that does the same thing my previous one did, just fixing a few special cases.
In particular, I noticed that the 'fpu_counter' situation was still rather confused in the last patch-series, and instead of having patch#1 fix up a part of it as part of re-organizing the FP state restore code, this just fixes the fpu_counter confusion separately. Especially since I wanted to clear the fpu_counter at fork time for new processes, just to make sure tht we never try to lazily switch to stale FPU state (re-used task struct pointer and all that).
So patch#1 is that fpu_counter work.
The two other patches are basically the same as last time, except with slightly updated commit logs and obviously updated for the fpu_counter fixes (it is, for example, wrong to update the counter at FPU preload time: if we do the lazy restore, there won't be any explicit preload, but the fpu_counter should still update - so that affected the preloading thing).
I've also done more testing of the series, although some of that shows funny effects: it's actually faster on my machine (and in my specialized tests) to switch between two processes where *one* uses the FPU with these lazy restore patches than it is to switch between two non-FPU users.
However, that seems to be due to some funny cache interaction: the profiles clearly show that '__switch_to()' itself is much faster if it doesn't have to work with any FPU state at all. But probably because of some random cache replacement detail, I then get more switches with the more expensive __switch_to when I only save the FP state (without ever restoring it).
Regardless, I'm pretty happy with this series, and I think I'll commit and push out at least the two first patches just because they fix real (albeit not very important) bugs. I feel like I'm going to do the third one too for 3.3, but I'm not sure yet. And comments welcome.
Linus
Linus Torvalds (3): i387: fix up some fpu_counter confusion i387: use 'restore_fpu_checking()' directly in task switching code i387: support lazy restore of FPU state
arch/x86/include/asm/i387.h | 53 +++++++++++++++++++++++++++---------- arch/x86/include/asm/processor.h | 3 +- arch/x86/kernel/cpu/common.c | 2 + arch/x86/kernel/process_32.c | 3 +- arch/x86/kernel/process_64.c | 3 +- arch/x86/kernel/traps.c | 40 +++++----------------------- 6 files changed, 54 insertions(+), 50 deletions(-)
-- 1.7.9.188.g12766.dirty
| |